File size: 6,539 Bytes
fb5206a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
"""
Evaluation metrics for transliteration:
  - Character Error Rate (CER)  β€” primary metric
  - Word Error Rate (WER)       β€” whole-word accuracy proxy
  - Top-1 Accuracy              β€” exact match
"""

import re
import logging
from typing import Dict, List, Optional, Tuple

import numpy as np
import evaluate                       # HuggingFace evaluate library
from jiwer import cer as jiwer_cer   # pip install jiwer

logger = logging.getLogger(__name__)

# ── CER / WER via jiwer ────────────────────────────────────────────────────────

def compute_cer(predictions: List[str], references: List[str]) -> float:
    """Character Error Rate β€” lower is better."""
    preds = [p.strip() for p in predictions]
    refs  = [r.strip() for r in references]
    return jiwer_cer(refs, preds)


def compute_wer(predictions: List[str], references: List[str]) -> float:
    """
    Word-level accuracy for single-word transliterations:
    WER = 1 - exact_match_rate in this context.
    """
    correct = sum(p.strip() == r.strip() for p, r in zip(predictions, references))
    return 1.0 - correct / len(references)


def compute_exact_accuracy(predictions: List[str], references: List[str]) -> float:
    """Fraction of predictions that exactly match the reference."""
    correct = sum(p.strip() == r.strip() for p, r in zip(predictions, references))
    return correct / len(references)


# ── Per-language breakdown ─────────────────────────────────────────────────────

def evaluate_per_language(
    predictions: List[str],
    references: List[str],
    source_texts: List[str],
) -> Dict[str, Dict[str, float]]:
    """
    Split predictions/references by language prefix and compute metrics
    per language.

    source_texts should contain the prefixed source, e.g. "__hi__ kitab".
    """
    from config import LANGUAGES, LANG_TOKEN

    buckets: Dict[str, Tuple[List[str], List[str]]] = {
        lang: ([], []) for lang in LANGUAGES
    }

    for pred, ref, src in zip(predictions, references, source_texts):
        for lang in LANGUAGES:
            if src.startswith(LANG_TOKEN[lang]):
                buckets[lang][0].append(pred)
                buckets[lang][1].append(ref)
                break

    results = {}
    for lang, (preds, refs) in buckets.items():
        if not preds:
            continue
        results[lang] = {
            "cer":      round(compute_cer(preds, refs), 4),
            "wer":      round(compute_wer(preds, refs), 4),
            "accuracy": round(compute_exact_accuracy(preds, refs), 4),
            "n_samples": len(preds),
        }

    return results


# ── HuggingFace Trainer-compatible compute_metrics ────────────────────────────

def build_compute_metrics(tokeniser):
    """
    Returns a compute_metrics function compatible with HuggingFace Seq2SeqTrainer.
    """

    def compute_metrics(eval_preds):
        pred_ids, label_ids = eval_preds

        # Decode predictions
        pred_ids = np.where(pred_ids != -100, pred_ids, tokeniser.pad_token_id)
        decoded_preds = tokeniser.batch_decode(pred_ids, skip_special_tokens=True)

        # Decode labels
        label_ids = np.where(label_ids != -100, label_ids, tokeniser.pad_token_id)
        decoded_labels = tokeniser.batch_decode(label_ids, skip_special_tokens=True)

        decoded_preds  = [p.strip() for p in decoded_preds]
        decoded_labels = [l.strip() for l in decoded_labels]

        cer_score = compute_cer(decoded_preds, decoded_labels)
        acc_score = compute_exact_accuracy(decoded_preds, decoded_labels)

        return {
            "cer":      round(cer_score, 4),
            "accuracy": round(acc_score, 4),
        }

    return compute_metrics


# ── Standalone evaluation on test set ─────────────────────────────────────────

def evaluate_model_on_test(
    model,
    tokeniser,
    test_dataset,
    batch_size: int = 256,
    num_beams: int = 4,
    device: str = "cpu",
) -> Dict:
    """
    Run beam-search decoding on the test set and return full metrics.

    Works with both HuggingFace models and CTranslate2 models
    (pass ct2_model=True to adjust decoding path).
    """
    import torch
    from torch.utils.data import DataLoader
    from datasets import Dataset
    from transformers import default_data_collator

    model.eval()
    model.to(device)

    all_preds, all_refs, all_sources = [], [], []

    loader = DataLoader(test_dataset, batch_size=batch_size, collate_fn=default_data_collator)

    with torch.no_grad():
        for batch in loader:
            input_ids      = batch["input_ids"].to(device)
            attention_mask = batch["attention_mask"].to(device)

            generated = model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                num_beams=num_beams,
                max_length=64,
            )

            preds   = tokeniser.batch_decode(generated, skip_special_tokens=True)
            labels  = batch["labels"]
            labels  = torch.where(labels != -100, labels, torch.tensor(tokeniser.pad_token_id))
            refs    = tokeniser.batch_decode(labels, skip_special_tokens=True)
            sources = tokeniser.batch_decode(input_ids, skip_special_tokens=False)

            all_preds.extend(preds)
            all_refs.extend(refs)
            all_sources.extend(sources)

    overall = {
        "overall_cer":      round(compute_cer(all_preds, all_refs), 4),
        "overall_accuracy": round(compute_exact_accuracy(all_preds, all_refs), 4),
    }
    per_lang = evaluate_per_language(all_preds, all_refs, all_sources)

    return {"overall": overall, "per_language": per_lang}


# ── CLI usage ──────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    # Quick sanity check
    preds = ["kitāb",  "namaste", "ধন্যবাদ"]
    refs  = ["kitaab", "namaste", "ধন্যবাদ"]
    print("CER:",      compute_cer(preds, refs))
    print("Accuracy:", compute_exact_accuracy(preds, refs))